Generating image captions in Polish – experimental study

摘要

The image captioning task focuses on generating sentences describing the content of a given image. Recently, some end-to-end solutions based on complex deep neural networks were proposed. However, most of them are producing captions in English, which is a syntactically relatively simple language. Application of captioning methods to more sophisticated languages, e.g., fusional, may cause problems due to the complex syntax. In the paper, we describe the initial results of applying the generative model to produce sentences in Polish, which is the fusional language. To validate the method, we used two datasets based on the Flickr8k database. The first is automatically translated Flickr8k, and the second one is the AIDe dataset consisting of 1000 pictures. The results were validated using BLEU, CIDEr, and BoW metrics.